Memorization Without Overfitting in Large Language Models

Created in September 06, 2023

2023   ·   llm   theory   ·   paper-review

This paper examines a counterintuitive phenomenon: LLMs can memorize specific training examples without suffering the generalization penalty typically associated with memorization in smaller models.

Key findings:

  • Memorization grows with model scale but does not necessarily correlate with reduced generalization
  • Rare and unique sequences are memorized disproportionately
  • Training dynamics: memorization of individual examples follows predictable curves during training
  • Implications for privacy: models can regurgitate verbatim training text, creating data extraction risks

The work raises fundamental questions about what “overfitting” means for foundation models trained on internet-scale data.



Enjoy Reading This Article?

Here are some more articles you might like to read next:

  • The Expressive Power of Transformers with Chain of Thought
  • Landscape of Thoughts — Visualizing Where LLM Reasoning Actually Goes
  • Magellan — Guided MCTS for Escaping the Gravity Wells of LLM Creativity
  • PriorZero — Injecting LLM Priors into MuZero-Style World Models at the MCTS Root
  • SuperThoughts — Reasoning Tokens in Superposition